v0.2.1

Fixed

• API_KEY-based model construction now works for every provider Pydantic-AI
  knows (openai-responses, groq, bedrock, …), not just a hand-maintained
  short list. build_model delegates the provider:name → Model-class mapping
  to Pydantic-AI's own infer_model, injecting the key via a provider_factory,
  so MODEL = "openai-responses:…" with an API_KEY no longer raises. A bare
  model name Pydantic-AI can map to a provider (e.g. claude-…) is accepted too.
• The drf-mcp toolset now sources each tool's schema from drf-mcp's own
  tools/list instead of re-deriving it from the input serializer alone. So the
  agent sees the full advertised inputSchema — a selector tool's
  filter / ordering / pagination arguments and the additionalProperties policy,
  not just the serializer's fields — matching the HTTP transport exactly.

Changed

• DEFAULT_SYSTEM_PROMPT gained gentle steering for two common failure modes:
  use a listing/search tool's arguments to find things by name and then act on
  the result (don't stop at the lookup), treat "open / go to / show me" as
  navigation, and always finish a turn with a reply or completed action.